def list_samplers(model, maximize=False):
node = model.dummy()
models.align(node, model)
samplers = get_samplers('dummy', node, maximize)
node.model.display()
print
for s in samplers:
print s
def fit_model(structure, data_matrix, old_root=None, gibbs_steps=200):
if old_root is None:
X = data_matrix.sample_latent_values(np.zeros((data_matrix.m, data_matrix.n)), 1.)
old_root = GaussianNode(X, 'scalar', 1.)
root = initialization.initialize(data_matrix, old_root, old_root.structure(), structure, num_iter=gibbs_steps)
model = models.get_model(structure, fixed_noise_variance=data_matrix.fixed_variance())
models.align(root, model)
dumb_samplers.sweep(data_matrix, root, num_iter=gibbs_steps)
dumb_samplers.sweep(data_matrix, root, maximize=True, num_iter=1)
return root
if __name__ == '__main__':
ensemble = list()
## Bow model requires the path to a pre-trained word2vect or GloVe vector space in binary format
#model = md.bow("/Users/fa/workspace/repos/_codes/MODELS/Rob/word2vec_100_6/vectorsW.bin")
#ensemble.append(md.bow("/home/ds/STS/GoogleNews-vectors-negative300.bin"))
## FeatureBased model is standalone and does not need any pre-trained or external resource
#ensemble.append(md.featureBased())
# Word align model based on Sultan et. al.
ensemble.append(md.align())
## Load some data for training (standard SICK dataset)
trainSet, devSet, testSet = load_data_SICK('../data/SICK/')
## Train a classifier using train and development subsets
classifier = train(ensemble, trainSet, devSet)
## Test the classifier on test data of the same type (coming from SICK)
test(ensemble, classifier, testSet).to_csv('../data/local/SICK-trained_SICK-test.csv')
## FileName to save the trained classifier for later use
fileName = '../data/local/SICK-Classifier.h5'
## VERSION THREE SAVE / LOAD (the only one that works)
classifier.save(fileName)
print len(allA)
print len(trainA) + len(devA) + len(testA)
print len(trainA), len(devA), len(testA)
return [trainA, trainB, trainS], [devA, devB, devS], [testA, testB, testS]
if __name__ == '__main__':
df = pd.DataFrame(columns=['bow', 'fb', 'aligner'])
ensemble = list()
## Load some data for training (standard SICK dataset)
trainSet, devSet, testSet = load_data_SICK('../data/SICK/')
bowm = md.bow("/home/ds/STS/GoogleNews-vectors-negative300.bin")
fbm = md.featureBased()
alignm = md.align()
ensemble.append(bowm)
ensemble.append(fbm)
ensemble.append(alignm)
classifiers = list()
## Train the different models in the ensemble using train and development subsets
for index, model in enumerate(ensemble):
classifier, df = train([model], trainSet, devSet, df)
classifiers.append(classifier)
print 'writing the final DataFrame'
df.to_csv('final' + str(index) + '.csv', sep='\t')
bow = pd.read_csv("final0.csv", sep='\t', engine='python')['bow']
